An Agent That Works While I Sleep

Public AI has a billing meter. Every request adds up, and a long agent run in the evening eats the plan I still need for tomorrow's actual work. Local AI does not.

Recently I used that difference on purpose. Before going to bed I handed my inference server the job I would not get done that evening: the missing CRUD endpoints in my current app, plus the tests that should have come with them. Clearing a field has to set it back to null, an organisation has to disappear along with its sites and appointments, and none of that was covered. Then I went to sleep.

After waking up

A branch with commits, the endpoints implemented on the server, and tests appended to each of the affected test files. A task list that started at one of five and ended empty. On llama.cpp with Qwen 3.8 27B.

The interesting part is the price. The box pulls around 650 watts under load. Eight hours of that is roughly five kilowatt hours, which at my tariff is about €1.50 for a night of work. No token counter, no rate limit, no "you have reached your weekly usage" in the middle of a refactoring. I can start something that finishes in an hour and something that runs for three days.

The downsides

A finished branch does not mean finished. I read the diff in the morning, because an agent that had eight hours undisturbed also had eight hours to be confidently wrong. Slow review is, so to speak, the tax on unattended work.

Two nights ago I woke up to a solution for a problem I described badly. The model had been extremely thorough in the wrong direction. That cost me a night.

Locally on my notebook, the model is unbearably slow when I wait in real time. If I just go to bed, it is bearable!

Open questions

For people on limited hardware the question is no longer "is an open model good enough for the task?" That question gets older every month. The useful question is: can this task wait until morning?

A lot of real work can. Migrations, test coverage, dependency updates, documentation, log analysis, the long tail of tickets nobody wants to hold in their head. Those tasks want patience, not speed.